Skip to content

Latest commit

 

History

History
108 lines (88 loc) · 4.77 KB

Install.md

File metadata and controls

108 lines (88 loc) · 4.77 KB

Installation and Dependencies of DOSA (gradatim)

Requirements

for RHEL

yum install python3.8 tmux python3-virtualenv
yum groupinstall 'Development Tools'
# for tvm
yum install llvm-toolset llvm-devel llvm-libs cmake

Python virtual environment

virtualenv venv -p /usr/bin/python3.8
. venv/bin/activate
pip install -r requirements.txt --no-dependencies

(Please note, the --no-dependencies is necessary to be able to install all required packages. All packages and their necessary dependencies are defined in the requirements.txt, but pip sometimes/often tries to solve it on its own and fails).

TVM

Dosa requires an installation of TVM, including python bindings. To be able to use the post-quantization feature of DOSA, our custom version of TVM is required: https://github.com/cloudFPGA/tvm-for-dosa

  1. Please follow the steps in TVM installation (some hints are also given in ./TVM-notes.md).
  2. add local tvm installation to environment. You have two alternatives:
    1. updated PYHTHONPATH as it is described here. (But we made the experience this is not always working with virutalenv or IDEs.)
    2. OR: add it as .pth to venv:
      • if virtualenvwrapper is installed: 
        source venv/bin/activate
add2virtualenv path/to/tvm/python path/to/tvm/vta/python
      • otherwise: 
        cp ./setup/_virtualenv_path_extensions.pth ./venv/lib/python3.8/site-packages/
        and update the absolute paths in the .pth file!

FPGA build tools

To generate FPGA binaries, the corresponding HLS, synthesis, and place and route tools must be installed. In case the Xilinx Vivado suite is used, ensure that the Y2K22 patch is installed.

Installation and Setup

git clone --recurse-submodules https://github.com/cloudFPGA/DOSA.git

DOSA runs purely in python, so besides the virtual environment (see above) only environment variables need to be set:

export TVM_LOG_DEBUG=0
# only necessary for builds on cloudFPGA
export DOSA_cFBuild1_used_dcps_path=/path/to/folder/with/shell_STATIC.dcp/
# sometimes necessary
export PYTHONPATH=.

The DOSA_cFBuild1_used_dcps_path is only required if IBM cloudFPGA is used as build platform. It is not required if DOSA is invoked with the --no-build flag. How to retreive the latest static Shells for the FPGAs (i.e. the .dcp files) is explained here.

Git submodules contained in DOSA

  • hls4ml in gradatim/backend/third_party_libs/
  • vhdl4cnn in gradatim/backend/third_party_libs/
  • cFDK in gradatim/backend/buildTools/lib/cFBuild1/
  • cFCreate in gradatim/backend/buildTools/lib/cFBuild1/
  • ZRLMPI in gradatim/backend/codeGen/templates/

Docker

DOSA, without the FPGA build tools, can also be build and run inside a docker container:

cd DOSA/doc
docker build -f Dockerfile -t dosa-build .

Afterwards, DOSA CLI can be started via:

docker run -it -v ./my_in_and_out_dir/:/scratch:Z -v ./folder_with_current_shell_STATIC/:/current_dcps/:Z dosa-build
# use DOSA CLI as described in usage
./gradatim -h
# for example, build one design with an example ONNX
./gradatim.sh onnx ./config/dosa_config_default.json ./examples/PTTCNN_int8.onnx ./examples/PTTCNN_meta.json /app/scratch/pttcnn/ --no-roofline

If building for the cloudFPGA target, the required Shell_STATIC.dcp files (as described above) should then be mounted to the /current_dcps/ folder inside the container. (Naturally, this disables the roofline diagram feature of DOSA, which requires a GUI.)

If DOSA should be called outside the docker container using the Pyro library, then the container should be started as follows:

docker run -it -v ./my_in_and_out_dir/:/scratch:Z -v ./folder_with_current_shell_STATIC/:/current_dcps/:Z dosa-build
# inside the container
python3 -m gradatim.pyro_server

This prints out a PYRO-URI that should be used by connected tools. Please note, that paths to e.g. the constraint file etc. then must be /scratch/path_to_file.json.